AI-enabled Prediction of Power Transformer Remaining Useful Life Using Dissolved Gas Analysis and Random Forest Regression

Abstract

Power transformers represent vital components within electrical power networks, and unexpected breakdowns can lead to significant financial losses and operational disruptions. Dissolved Gas Analysis remains a widely adopted technique for monitoring transformer health, yet most diagnostic methods and machine learning applications concentrate on fault identification rather than ongoing prognostic evaluation. Persistent challenges include the nonlinear nature of fault development, correlations among gas variables, and limited transparency in model outputs. This research introduces a data-driven framework designed to estimate transformer Remaining Useful Life through Dissolved Gas Analysis combined with an optimised Random Forest regression approach. The framework is validated using a publicly accessible dataset containing 2,100 labelled instances of dissolved gas readings, obtained from the Kaggle repository on transformer faults and Remaining Useful Life prediction. Statistical descriptors of hydrogen, carbon monoxide, acetylene, and ethylene are employed after targeted feature selection to address multicollinearity. Relative to traditional Dissolved Gas Analysis interpretation and baseline machine learning techniques, the proposed model achieves higher predictive accuracy, lower estimation error, and enhanced robustness. An analysis of feature importance further differentiates the framework by offering clear insight into the gas parameters most strongly associated with Remaining Useful Life. The findings substantiate the effectiveness of Random Forest regression in delivering dependable, interpretable, and practically applicable predictions of transformer service life.